Developing machine learning and data mining techniques for fuel loss detection at service stations

This research aims to develop effective machine-learning algorithms for detecting unexpected volume changes and identifying their potential error sources in underground storage tanks (USTs) in service stations. Detecting fuel losses has been one of the significant challenges all over the world. Many site operators utilise sensors placed at strategic locations around their service station to collect data and identify potential fuel loss issues. However, the fluctuating weather conditions across Australia, the differences in fuel composition, and the inaccuracies in measurement equipment might cause fuel loss and gain in the tank, and it is very difficult to identify these unexpected volume changes. Leaks, one of the most essential and crucial types of fuel loss, can have a devastating impact on the environment and pose significant health risks to human lives. Existing leak detection systems based on Statistical Inventory Reconciliation (SIR) analysis mainly use rudimentary statistical models. Thus, they may be slow to identify unexpected volume changes and fail to capture small loss behaviour. In summary, there is a pressing need for the establishment of better loss detection systems. Additionally, investigating all the loss-related alarms can waste time, resources, and money, whereas ignoring critical alarms might compromise site compliance and cause environmental damage. Thus, identifying the potential underlying error source of the alarm and speeding up the alarm investigation process for the technicians is necessary. In this study, the fuel loss problem is approached as an anomaly detection task on multivariate time series data. A novel methodology that integrates algorithmic and systemic design principles is presented to identify and diagnose anomalies by triggering alarms and pinpointing potential sources of errors, with the ultimate goal of comprehending the underlying causes of the detected anomalies. The architecture of this system design consists of four main stages: (1) Data Collection, (2) Data Pre-processing, (3) Anomaly Detection, and (4) Anomaly Diagnosis and Explainability. The study utilises three primary sensor data sources that exhibit various data problems and inaccuracies. As such, the Data Pre-processing stage includes three preliminary data pre-processing phases and two additional steps to prepare the data for modelling. Our results show that the provided comprehensive data pre-processing framework enhances the quality of sensor data and enables more effective anomaly detection. Based on a thorough literature review, the Anomaly Detection stage employs algorithms such as the statistical model “TUBE” algorithm, which is specifically designed for leak detection to compare its performance with current industry practices. On the other hand, we evaluated some of the most commonly utilised anomaly detection algorithms, such as the traditional machine learning model “OC-SVM” and the deep learning model “LSTM-AE”. This study introduces a new methodology for identifying loss behaviour as a collective anomaly detection issue. By utilising multivariate deep learning algorithms, the study demonstrates that this approach surpasses current industry methods in terms of detecting fuel loss at lower fuel leak rates. This finding suggests that the proposed methodology can detect fuel loss behaviour at an earlier stage than current industry practices due to the gradually incremental nature of fuel loss. The Anomaly Diagnosis and Explainability stage aims to determine the alarm error source categories for the anomaly periods and to reduce alarm resolution time for industry practitioners. In this stage, the focus is on explaining the alarms generated by the LSTM-AE model through feature importance-based, distance-based and SHAP-based explainability models. The results of this project have the potential to increase productivity and provide effective techniques for accurately detecting fuel losses in USTs and determining the underlying causes of fuel loss alarms.